Improved BLS based transformer fault diagnosis considering imbalanced samples

Abstract

At present, the traditional machine learning diagnosis method has some problems, such as the imbalance of various fault types and the great difference of recognition effect between different fault types. In order to solve the problem of low accuracy of traditional machine learning diagnosis methods, a multilevel fault diagnosis model of transformer based on hierarchical classification, generalized learning and ensemble learning is established. According to the imbalanced degree of each category sample, the method established the corresponding classifier, and carried on the diagnosis step by step. At the first level, the broad learning system was selected to extract three generalized feature labels—normal, discharge and overheat, which were fused with the original parameter input to guide the nine more detailed state types. In the second-level classifier, EasyEnsemble learning method is used to generate multiple data-balanced training subsets through under-sampling, fully balancing the fault information of most classes and a few classes, and then synthesized the final classifier through parallel training sub-classifiers, avoiding the problem of missing data information due to under-sampling. Experimental results show that, compared with the traditional diagnosis methods, the proposed method improved the generalization characteristics of a few types of faults, improved the overall accuracy rate, and had a more accurate and balanced fault diagnosis effect of power transformer.

Keywords

• Imbalanced samples
• Hierarchical classification
• Broad learning system
• Ensemble learning